Macromolecular Modeling Blog ™

About a month ago Yose Widjaja posted a question to the pdb-l mailing list:

” Suppose you have two pdb files of two proteins you suspect interact with each other. What existing approaches exist out there that can tell you whether these two proteins interact, based on the structural information alone? “

His question was quickly answered by prominent figures in the structural bioinformatics world, as Kevin Karplus

” Although there are plenty of people working on this problem, I’d have to say that at the moment it is not possible to tell whether two proteins interact by looking at their structures (unless they are members of a pair of families known to interact routinely). For some interesting families (like kinases) it is very difficult to predict what proteins will interact even if you have homologs that are co-crystallized. Currently, structural assistance in determining protein-protein interaction is marginal at best. If you can’t do it from sequence alone, knowing the structures rarely helps. Note: I’d be delighted to find out that I’m wrong, and that there has been substantial progress in determining protein-protein interactions from structural information. If you want to make that claim, give me some citations of papers to read to catch up with the field! “

and Julian Mintseris

“…one could argue that an algorithmic determination of whether two proteins interact is pretty much impossible.  It’s not really a binary problem – the best that an algorithm could do is show that for a range of proteins complexes, it can give a reasonable correlation of Kd values, but the significance of that is also unclear given how much Kd can vary based on experimental conditions, etc. “

This conversation inspired us to check out what is the state of predicting protein-protein interaction based on structural knowledge, or the “specificity prediction” problem as it is sometimes termed. It seems that in point of fact, this is a very difficult problem, and there is little advancement in the field. Indeed, we failed to find any general approach, that given two structures, predict whether they bind. That is, without any other information, specifically without relaying on homologs. 

Perhaps the most general scheme we found is: M-TASSER: an algorithm for protein quaternary structure prediction. M-TASSER is threading the sequences of the protein pair onto available dimeric templates. On a dataset of 207 targets predicted to interact as dimers, M-TASSER achieved a true positive rate of 68% and a false positive rate of 17%. 

The other two prominent groups we identify that tackle with this problem have each focused on a specific family of interactions, and for that family show nice results:

Amy Keating is developing the Structure-based Prediction of bZIP Partnering Specificity. Using a machine-learning approach which combines a variety of functions based on physics, learned empirical weights or experimental coupling energies, helix propensities and residue-residue interactions, they were able to correctly order the stabilities of over 6000 pairs of bZippers complexes with greater than 90% accuracy. The final model is physically interpretable, and suggests specific pairs of residues that are important for bZIP interaction specificity.

Luis Serrano is focusing in constructing A genome-wide Ras-effector interaction network. Using structural information and protein design tools (i.e. FoldX), they predict the interactions between 20 Ras subfamily proteins with 50 putative Ras binding domains. The sequences of these were modeled onto existing templates, and the energy for the models was used as a predictor. Pull-down experiments were used to validate some of the predictions, and literature mining as well. The method shows approximately 80% accuracy for distinguishing between binders and non-binders.

Do you know of any other such specificity prediction works? What other systems have enough structural information to allow for such approaches? tell us in the comments.

Be Sociable, Share!

•